a census-based approach to spanish language maintenance

• Census nuts/bolts
• New Mexico & the US
• Some macro-exploration
• A simple model
• Some final notes
• References

Census nuts/bolts

Language data for the Hispanic population live in census table B16006, which provides summary counts of speakers by language spoken at home and level of English proficiency for the population over 5.

To investigate variables comprising a particular census table, we can use the load_variables function from the tidycensus package.

spanLang <- load_variables(2015, "acs5", cache = FALSE) %>%
  filter(grepl("B16006",name)==TRUE & !grepl("Margin|other",label))%>%
  mutate(name=gsub("E","",name),label=gsub(":","",label))%>%
  select(-concept)

The table below summarizes the variables included in census table B16006. A third column, speakerType, is added, which reflects a re-categorization of census variables per our purposes here.

Approximating Spanish language maintenance

As the table details, we can use census variables to identify Spanish speaking Hispanics as either bilingual English-Spanish speakers or monolingual Spanish speakers. Hispanics that “speak only English” at home are assumed here to be monolingual English speakers. Rates of each speaker type can then be calculated using the Hispanic population over 5 as denominator.

Based on these distributions, we follow Bills (1989) in approximating the language maintenance rate for a given community as the sum of rates of monolingual Spanish speakers and bilingual speakers.

In addition to maintenance rates, we characterize each community/geography in terms of its overall population and its Hispanic population density, ie, the percentage of a community that identifies as Hispanic. This very simple demographic profile can be derived from variables included in census tables B01001I and B01001.

Gathering ACS data

To gather these two sets of data for the most recent 5-year ACS estimates (2011-15), we use the the tidycensus package. Our query is comprised of seven variables from three tables:

langVars <- c("B16006_001", 
              "B16006_002", 
              "B16006_003", 
              "B16006_004", 
              "B16006_005", 
              "B16006_006", 
              "B16006_007", 
              "B01001I_001", 
              "B01001_001")

Relevant geographies include all metropolitan statistical areas (MSAs), all US States, and the US.

geos <- c("us", 
          "metropolitan statistical area/micropolitan statistical area", 
          "state")

We then apply the tidycensus::get_acs function across each geography to obtain our data set.

summaryData <- lapply(geos, function (x) {
  tidycensus::get_acs (geography= x, 
                       variables=langVars, 
                       output="wide", 
                       year=2015)}) %>%
  bind_rows() 

Lastly, we perform some data cleaning & transformation processes.

summaryData<- summaryData %>%
  rename(HispPopOver5=B16006_001E, HispPop=B01001I_001E, TotalPop=B01001_001E)%>%
  group_by(NAME)%>%
  mutate(BiLing = round(sum(B16006_004E, B16006_005E)/ HispPopOver5*100,1),
         EngMono = round(B16006_002E/ HispPopOver5*100,1),
         SpanMono = round(sum(B16006_006E, B16006_007E)/ HispPopOver5*100,1), 
         PerHisp = round(HispPop/ TotalPop*100,1))%>%
  select(NAME, GEOID, HispPopOver5, TotalPop, HispPop, PerHisp, EngMono, BiLing, SpanMono)%>%
  ungroup()%>%
  mutate(NAME = gsub("-.*(,)","\\1",NAME) %>% gsub(" Metro Area","",.))

New Mexico & the US

So, we first take a comparative look at rates of Spanish language maintenance in New Mexico and some other MSAs rich in Hispanic heritage. Geographies in this subset, then, include the US, the state of New Mexico, all MSAs in New Mexico, and ten additional MSAs from around the country.

geoSample <- c("1","35","10740", 
               "29740", "22140", 
               "42140", "26420", 
               "12060", "38060", 
               "16980", "19740", 
               "31080", "12420", 
               "27260", "33100", 
               "46060")

The table below summarizes these results. The total population (TotalPop) is in 1K units. PerHisp = percentage of the population that identifies as Hispanic (ie, Hispanic population density), EngMono = percentage of Hispanic population that speaks only English, BiLing = percentage of Hispanic population that speaks both Spanish and English, and SpanMono = percentage of Hispanic population that speaks only Spanish.

Columns can be sorted by clicking on the column name.

summaryData %>%
  filter(GEOID %in% geoSample) %>%
  select(-HispPopOver5, -GEOID, -HispPop)%>%
  mutate(TotalPop=round(TotalPop/ 1000))%>%
  arrange(desc(PerHisp))%>%
  DT::datatable(extensions = 'FixedColumns',
                options = list(pageLength = length(geoSample),
                               dom = 't',
                               scrollX = TRUE,
                               fixedColumns = list(leftColumns = 1)), 
                rownames = FALSE,
                width="100%", 
                escape=FALSE) %>%
  DT::formatStyle(names(summaryData[6:9]),
    background = styleColorBar(range(summaryData[6:9]), "cornflowerblue"),
    backgroundSize = '80% 70%',
    backgroundRepeat = 'no-repeat',
    backgroundPosition = 'right') %>%
  DT::formatStyle(c(1:6),Size = '85%')

As can be noted, there is a considerable amount of variation in the distribution of speaker-types across geographies. A less interactive, but more comprehensive display of this variation is presented in the figure below. The plot is sorted by rates of language maintenance (ie, SpanMono + BiLing). The percentage of Hispanics that are monolingual English speakers is transformed to a negative value to hack together a likert-like plot in a simple fashion.

summaryData %>%
  filter(GEOID %in% geoSample) %>%
  select(NAME, EngMono, BiLing, SpanMono)%>%
  mutate(rank=rank(EngMono), EngMono = -EngMono)%>%
  gather(key=speakerType, value=pct ,EngMono ,BiLing, SpanMono)%>%
  mutate(speakerType = factor (speakerType,levels = c("EngMono","SpanMono","BiLing")))%>%
  ggplot(aes(x=reorder(NAME,-rank), y=pct, fill=speakerType))+
    geom_bar(stat="identity",width=.75) +
    scale_fill_manual(values = c("#bdc9e1","#0570b0", "#74a9cf" ))+
    labs(x="",y="Percentage")+
    coord_flip()+
    theme_fivethirtyeight() +
    theme(legend.position = "bottom",legend.title = element_blank(),
          plot.title = element_text(size=15))+ 
      labs(title = "Language maintenance rates by geography")

So, some quick observations from the table and figure.

• Relative to other MSAs in the US, Spanish language maintenance rates in New Mexican MSAs are some of the lowest, despite some of the highest rates of Hispanic population density in the country. Las Cruces is an exception.
• In comparison to MSAs with higher rates of Spanish language maintenance, MSAs in New Mexico are decidedly less populous. While Hispanics comprise only 10% of Atlanta’s total population, for example, roughly 80% of Hispanics in Atlanta speak Spanish.
• Spanish monolinguals seem to contribute more to overall maintenance rates in MSAs outside of New Mexico.

Some macro-exploration

Next, we consider variation in maintenance rates as a function of geographic location, Hispanic poulation density, and overall poulation for all MSAs in the (continental) US.

Geospatial variation

MSAs are geographical areas (comprised of counties) that often span multiple states. As polygons, they do not make for fantastic maps as they do not provide coverage of the entire US. A visually cleaner approach is to represent each MSA as a single point (or centroid) instead. This additionaly allows us to add some regional geographic data (eg, US divisions) to our demographic-linguistic profile of each MSA via a spatial join.

Using the tigris package, we import a US Divisions polygon as an sf spatial object:

library(tigris)
options(tigris_class = "sf",tigris_use_cache = TRUE)
library(sf)

divs <- tigris::divisions(cb=TRUE)%>%
  st_transform(crs = "+init=epsg:4326")%>%
  select(NAME)

US divisions include:

## [1] "New England"        "Middle Atlantic"    "East North Central"
## [4] "West North Central" "South Atlantic"     "Mountain"          
## [7] "Pacific"            "East South Central" "West South Central"

Next, we import a MSA polygon, get the centroids of each polygon, and spatially join the US division polygon to the centroids. All of this can be accomplished in the following pipe with the help of the sf package.

metros <- tigris::core_based_statistical_areas(cb=TRUE)%>% #import poly
  rename(msa_name=NAME)%>%
  st_transform(crs = "+init=epsg:4326")%>%
  st_centroid() %>% #get poly centroids
  sf::st_join(divs)%>% #spatial join with us divisions
  select(GEOID,NAME)%>%
  na.omit

Limit analysis to MSAs in the continental US, and then join our new MSA-division point geometry:

msas_w_divs <- summaryData %>%
  filter(!grepl("PR|Micro|HI|AK",NAME) & nchar(GEOID)>2)%>%
  left_join(metros %>% rename(div=NAME))%>%
  mutate(speak_span = BiLing+SpanMono)%>%
  st_as_sf()

Finally, we map rates of Spanish language maintenance by MSA using the leaflet package. In the map below, rates have been transformed into nine quantiles to get a clearer look at variation. The reddest points reflect rates in the lowest 11% of the distribution; the bluest points reflect rates in the highest 11% of the distribution. Yellow points reflect median rates of Spanish language health. Polygons are US divisions.

Hover over the legend to see how quantiles translate to underlying rates of Spanish language maintenance.

library(leaflet)
library(widgetframe)

pal <- colorQuantile(palette = 'RdYlBu', domain = msas_w_divs$speak_span, n = 9)

mp <- leaflet(divs,width="100%",height='400') %>%
      setView(lng = -98.35, lat = 39.5, zoom = 4) %>%
      addProviderTiles ("CartoDB.Positron", 
                        options = providerTileOptions(minZoom = 4, maxZoom = 5))%>%
      addPolygons(color="gray",
                  stroke = TRUE,
                  weight=1,
                  fillOpacity = 0.1) %>%
      addCircleMarkers(data=msas_w_divs, 
                       radius=3.5, 
                       fillColor=~pal(speak_span),
                       stroke = FALSE, fillOpacity = 1,
                       label=~paste(div,' - ',NAME))%>% 
      addLegend("bottomright", 
                pal = pal, 
                values = ~ msas_w_divs$speak_span, 
                title = "Quantiles", 
                opacity = 1)

The map illustrates higher rates of Spanish language maintenance in coastal and border states. Per our previous observation, MSAs in the Southwest (with a few exceptions) are largely in the red (ie, below-median) when it comes to maintenance rates.

Geo-demographic variation

Here we explore the independent relationships between total population and maintenance rates, on one hand, and Hispanic population density and maintenance rates, on the other hand.

In the scatter plots below, both independent variables have been log-transformed to smooth out their distributions some. To get a better sense of geographical variation, MSA names are included in plots; colors reflect US division. Note that not all data points are included in the plots to avoid overlap and aid in readability.

msas_w_divs%>%
  data.frame()%>%
  mutate_at(vars(PerHisp,TotalPop), funs(log)) %>%
  gather(key= langVar,value = val,c(PerHisp,TotalPop))%>%
  ggplot(aes(x=val, y=BiLing+SpanMono)) + 
  geom_smooth(method="loess", se=T) +  
  geom_text(aes(label=NAME,col=div), 
            size=2.5, 
            check_overlap = TRUE,
            hjust = "inward")+
  facet_wrap(~langVar, ncol=1,scales = "free_x") +
  scale_colour_stata() + theme_fivethirtyeight() +
  theme(legend.position = "bottom") 

Plots both suggest a fairly strong relationship between independent variable and Spanish language maintenance.

A simple model

For a more comprehensive perspective on these relationships, we build a simple model with maintenance rates as the dependent variable, and Hispanic population density, total population, and US Division as independent variables.

summary(lm(SpanMono + BiLing ~ log(PerHisp) + log(TotalPop) + div,
           data=msas_w_divs))
## 
## Call:
## lm(formula = SpanMono + BiLing ~ log(PerHisp) + log(TotalPop) + 
##     div, data = msas_w_divs)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -32.479  -6.462   0.780   7.929  27.598 
## 
## Coefficients:
##                       Estimate Std. Error t value Pr(>|t|)    
## (Intercept)            10.7755     6.9322   1.554 0.120948    
## log(PerHisp)           11.4482     0.7671  14.923  < 2e-16 ***
## log(TotalPop)           1.9372     0.5613   3.451 0.000623 ***
## divEast South Central  15.8155     2.4783   6.382 5.31e-10 ***
## divMiddle Atlantic      2.3337     2.3602   0.989 0.323419    
## divMountain           -15.0990     2.5170  -5.999 4.78e-09 ***
## divNew England          4.7034     3.1821   1.478 0.140243    
## divPacific             -5.6994     2.4773  -2.301 0.021974 *  
## divSouth Atlantic      10.7062     1.8884   5.670 2.91e-08 ***
## divWest North Central  -2.0168     2.3762  -0.849 0.396578    
## divWest South Central  -1.2937     2.3720  -0.545 0.585803    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 10.98 on 366 degrees of freedom
## Multiple R-squared:  0.5537, Adjusted R-squared:  0.5415 
## F-statistic: 45.41 on 10 and 366 DF,  p-value: < 2.2e-16

Model results

So, results of the model suggest that rates of Spanish language maintenance are higher in:

• densely Hispanic MSAs,
• highly populated MSAs, and
• the South.

Additionally, results suggest that Hispanics are less likely to speak Spanish in MSAs located in the Mountain division.

Some cursory explanations

In terms of accounting for the independent effects of the two population-based metrics, on one hand, densely Hispanic MSAs imply a more pervasive Hispanic culture; on the other hand, highly populated MSAs imply an overall culture of diversity. In theory, each account provides a distinct “mechanism” for continued language use.

Higher rates of Spanish language maintenance in the South likely reflect more recent immigration patterns. In contrast, lower rates in the Mountain division (including the Southwest) could reflect a combination of higher levels of acculturation among these Hispanic populations and lower immigration rates (as suggested in Bills 1989).

Some final notes

So, a quick and very much exploratory investigation into Spanish language maintenance in the US, using census data as an imperfect proxy. Grains of salt abound for sure. Factors influencing whether or not speakers continue to use a minority language are many and complex, and are only superficially addressed here.

Many of these patterns (at least in the Southwest) have been observed previously; the goal of this post was simply to demonstrate an innovative methodology based in previous research using a reproducible example.

References